Tag
Introduces CODA, a GPU kernel abstraction that expresses Transformer operations as GEMM-plus-epilogue programs to reduce data movement, covering nearly all non-attention computation in a Transformer block.